Search and rescue operations in wildfires using unmanned aerial vehicles: A multi-agent deep reinforcement learning approach
Maxime Collignon, Adolfo Perrusquía, Antonios Tsourdos, Weisi Guo
Abstract
Wildfires pose major challenges to natural ecosystems and smart living due to its destructive nature. Unmanned Aerial vehicles (UAVs) or drones have been used to support fire fighter in identifying vulnerable areas and the detection of people that need assistance. Most of the current solutions use path planning approaches under simple and deterministic environments that fail to model the dynamic nature of fire. Furthermore, the localisation of victims is assumed to be known which is unrealistic in disaster-like scenarios. To alleviate this issue, this paper proposes a novel search and rescue (SAR) application using drones. A multi-agent deep Q-network is designed to train a fleet of UAVs to search for people and evacuate them in a wildfire scenario. A realistic forest environment is designed that considers variations in vegetation and fire propagation. This helps to challenge RL algorithms to be more adaptive to changes in the environment due to the evolution of fire. Extensive simulation experiments are conducted to show the advantages and effectiveness of the proposed approach.